19 lines
3.0 KiB
Markdown
19 lines
3.0 KiB
Markdown
# H2-S2 actually started; first-epoch cost gate PASS
|
||
|
||
The sole service `meltmind-h2-s2-20261117` runs coordinator701941, guards701953/701954/701955/701956, and fold1/2/3/4 workers702009/702019/702016/702018. Actual-start evidence records four independent three-poll idle checks, completed learning-only normalization, real epoch1 optimizer steps and matching source pins. This is actual computation, not just a service-start acknowledgment.
|
||
|
||
Root `/home/frank14f/MeltMind-runs/phase2-h2-s2-2026-11-17`; implementation1f57bad66eb062d4f2269347532321929f873717; contract76e1976bed88fc390812238c3d80693446daea3840dc009ac7d8870ddc43a31d. Sixteen CPU synthetic tests passed; original failed fixture log retained. All29source pins are unchanged. Remote implementation508ea39d00c5cd51847e8b1aaa0c1cce614a8672 is ref-verified; independent HTTP readback of worker,guard,queue,contract and execution document matched all5hashes.
|
||
|
||
| Fold | Learning parts | First complete training epoch, s | Projected new fold GPU-s | Fixed fold ceiling |
|
||
|---|---:|---:|---:|---:|
|
||
| 1 | 374 | 102.782502789 | 26080.5329920 | 30685.4462728 |
|
||
| 2 | 366 | 107.558574978 | 27244.2203684 | 30685.4462728 |
|
||
| 3 | 291 | 89.451310205 | 22891.3378233 | 30685.4462728 |
|
||
| 4 | 350 | 107.500166093 | 27208.0728030 | 30685.4462728 |
|
||
|
||
The exact first-epoch projection is124682.37889553462GPU-s including prior18858.21490871499 once, all projected training/setup/waits,1.25safety and2400future-prediction reserve; it passes144000. All jobs also pass their individual ceilings. The original queue automatically released the barrier and continued. First-epoch audit at1794925327 saw all four inepoch2, cumulative optimizer steps664/651/566/621,9expectedprocesses,sourcepinsPASS and no global stop. That snapshot, four immutable001epoch receipts, normalization, fold identity and first-cost gate are fixed ine767d5777c4d3bc0a16018113ac95a00f24938c7. This is not a200epoch training-completion or pooled-score result.
|
||
|
||
The same queue refreshes its timing projection at common ten-epoch milestones and stops on failure or any budget/identity/resource violation. No retries, source/threshold changes, fold0 rerun or data-sufficiency launch. Training stops after all four PASS/checkpoint freeze, at TRAIN_COMPLETE_POSTFIT_PENDING. A separately bound pooled scorer needs a new actual Assist pre-score semantic PASS; S1's receipt must not be reused. Confirmation remains552/204,component means equally weighted,2000replicates/seed2026111408,pooled field H2−A CI upper<0; B13/last/tau report-only.
|
||
|
||
The original gate-outcome coordinator666520 is unchanged. The same audit read321terminal/2active/170pending and64549.80548118CPU-s, with original source pinsPASS. Its live cap is still72000total/68400geometry. The approved95000final ceiling permits one labeled continuation only after the original queue actually stops at its budget and exits, with all terminal hashes and old CPU counted once; original failures remain. No gate continuation or new review identity was launched here.
|